The Paper2Agent system converts academic papers into interactive AI agents capable of executing research methods. By analyzing text, code, datasets, and workflows, the system creates an MCP-based server that allows language models to run specific scientific procedures without manual environment setup.
During testing, the system created 22 tools for the AlphaGenome model in 45 minutes at a cost of $14. The agent achieved 98.7% accuracy on 15 training tasks and 100% on 15 new tasks. For the Scanpy package, it generated 7 tools in 45 minutes for $13, successfully performing data processing and cell type identification.
In a broader trial involving 100 biology papers, Paper2Agent successfully converted 74 publications into agents, producing 599 tools with 593 passing validation. Failures were primarily due to missing code, data, or dependency issues. The researchers also demonstrated the system's applicability to fields such as statistics, econometrics, game theory, and astrophysics.
The developers note that Paper2Agent does not replace human verification or guarantee the validity of research. While many papers remain incompatible due to poor documentation or code quality, the system enables researchers to apply methods directly to their own data or reproduce analyses without manual implementation.
